Bilingually-constrained Synthetic Data for Implicit Discourse Relation Recognition

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2306–2312, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Bilingually-constrained Synthetic Data for Implicit Discourse Relation Recognition Changxing Wu1,2, Xiaodong Shi1,2∗, Yidong Cheng1,2, Yanzhou Huang1,2, Jinsong Su3 Fujian Key Lab of the Brain-like Intelligent Systems, Xiamen University, China1 School of Information Science and Technology, Xiamen University, China2 School of Software, Xiamen University, China3 {wcxnlp, huangyanzhou163}@163.com {mandel, ydchen, jssu}@xmu.edu.cn Abstract To alleviate the shortage of labeled data, we propose to use bilingually-constrained syn- thetic implicit data for implicit discourse re- lation recognition. These data are extracted from a bilingual sentence-aligned corpus ac- cording to the implicit/explicit mismatch be- tween different languages. Incorporating these data via a multi-task neural network model achieves significant improvements over baselines, on both the English PDTB and Chi- nese CDTB data sets. 1 Introduction Discovering the discourse relation between two sen- tences is crucial to understanding the meaning of a coherent text, and also beneficial to many down- stream NLP applications, such as question answer- ing and machine translation. Implicit discourse re- lation recognition (DRRimp) remains a challenging task due to the absence of strong surface clues like discourse connectives (e.g. but). Most work re- sorts to large amounts of manually designed features (Soricut and Marcu, 2003; Pitler et al., 2009; Lin et al., 2009; Louis et al., 2010; Rutherford and Xue, 2014), or distributed features learned via neural net- work models (Braud and Denis, 2015; Zhang et al., 2015; Ji and Eisenstein, 2015). The above methods usually suffer from limited labeled data. Marcu and Echihabi (2002) attempt to create la- beled implicit data automatically by removing con- nectives from explicit instances, as additional train- ing data. These data are usually called as syn- ∗Corresponding author. thetic implicit data (hereafter SynData). How- ever, Sporleder and Lascarides (2008) argue that SynData has two drawbacks: 1) meaning shifts in some cases when removing connectives, and 2) a different word distribution with the real implicit data. They also show that using SynData directly degrades the performance. Recent work seeks to de- rive valuable information from SynData while fil- tering noise, via domain adaptation (Braud and De- nis, 2014; Ji et al., 2015), classifying connectives (Rutherford and Xue, 2015) or multi-task learning (Lan et al., 2013; Liu et al., 2016), and shows promising results. ch: [社会 认为 有 青少年 问题,]Arg1 en: [society reckons the existence of youth problems, ]Arg1 implicit=但是 [很多 青少年 认为 自己 没 问题.]Arg2 but [many young people do not think there is anything wrong with them.]Arg2 society reckon existence youth problems, but many young people think themselves no problems. Figure 1: An example illustrating the implicit/explicit mis- match between Chinese (ch) and English (en). A Chinese im- plicit instance is translated into an English explicit one. In the PDTB, a discourse instance is defined as a connective (e.g. but) taking two arguments (Arg1 and Arg2). Different from previous work, we propose to con- struct bilingually-constrained synthetic implicit data (called BiSynData) for DRRimp, which …

Content truncated to 500 words